基于DCE-MRI瘤内及瘤周影像组学联合TIC分型及Ki-67预测乳腺癌患者HER-2表达

Prediction of HER-2 expression in breast cancer patients based on DCE-MRI intratumor and peritumoral imaging combined with TIC typing and Ki-67

目的 探讨基于动态对比增强MRI(dynamic contrast enhancement MRI, DCE-MRI)瘤内及瘤周的影像组学模型联合临床、影像学指标预测乳腺癌患者人表皮生长因子受体2(human epidermal growth factor receptor 2, HER-2)表达状态的价值。材料与方法 回顾性收集2018年6月至2022年9月经病理证实为乳腺癌的患者病例资料272例,其中HER-2阳性139例,阴性133例,所有病例均在治疗前进行DCE-MRI检查。采取7:3的比例随机分为训练集和验证集。在训练集中使用皮尔森相关系数、递归特征消除法、逻辑回归对瘤内及瘤周影像组学数据进行降维及模型构建;利用多因素logistic回归筛选临床及影像学资料中的独立危险因素,以此构建临床模型;最终以瘤内、瘤周及临床特征构建联合模型。采用受试者工作特征曲线下面积(area under the curve, AUC)评价模型的效能,应用决策曲线分析(decision curve analysis, DCA)评估模型的临床价值。结果 临床模型、瘤内模型、瘤周模型、瘤内+瘤周模型及联合模型在训练集的AUC分别为0.736、0.784、0.806、0.831、0.854,准确度分别为69.5%、70.5%、75.8%、73.7%、76.8%,敏感度分别为87.6%、53.6%、71.1%、62.9%、72.2%,特异度分别为50.5%、88.2%、80.6%、84.9%、81.7%;在验证集中的AUC分别为0.731、0.724、0.713、0.780、0.799,准确度分别为73.2%、70.7%、68.3%、73.1%、78.0%,敏感度分别为76.2%、61.9%、88.1%、76.2%、78.6%,特异度分别为70.0%、80.0%、47.5%、70.0%、77.5%。经DeLong检验,训练集中联合模型与临床模型、瘤内模型、瘤周模型之间差异有统计学意义(Z=3.660、2.791、2.201,P=0.0003、0.005、0.028),联合模型与瘤内+瘤周模型之间差异无统计学意义(Z=1.583,P=0.114)。结果表明在训练集和验证集中联合模型对预测HER-2的状态优于临床模型、瘤内模型、瘤周模型及瘤内+瘤周模型。DCA显示在训练集中风险阈值在13%~60%时联合模型较临床模型、瘤内模型、瘤周模型及瘤内+瘤周模型具有更高的临床价值。结论 基于DCE-MRI瘤内及瘤周影像组学联合临床、影像学特征构建的联合模型能够较好地预测乳腺癌患者HER-2的表达状态。

Objective:To investigate the value of dynamic contrast enhancement MRI(DCE-MRI)based intratumoral and peritumoral radiomics models in combination with clinical and imaging indicators to predict the expression status of human epidermal growth factor receptor 2(HER-2)in breast cancer patients.Materials and Methods:A total of 272 patients'information with pathologically confirmed breast cancer from June 2018 to September 2022 were retrospectively collected,including 139 patients with positive HER-2 and 133 patients with negative HER-2.All cases underwent DCE-MRI examination before treatment.All 272 patients were divided into training set and validation set with a ratio of 7:3 by complete randomization method.In the training set Pearson correlation coefficients,recursive feature elimination and logistic regression were used to perform dimensionality reduction and model construction of intratumoral and peritumoral radiomics data.Multivariate logistic regression was used to screen the independent risk factors in clinical and imaging data,so as to construct the clinical model.Finally,the combined model was constructed by using intratumoral,peritumoral and clinical features.Area under the curve(AUC)was used to evaluate the efficacy of the model,and decision curve analysis(DCA)was used to evaluate the clinical value of the model.Results:The AUC of clinical model,intratumoral model,peritumoral model,intratumoral+peritumoral model and combined model in the training set were 0.736,0.784,0.806,0.831,0.854,and the accuracy was 69.5%,70.5%,75.8%,73.7%,76.8%,respectively.The sensitivity was 87.6%,53.6%,71.1%,62.9%,72.2%,and the specificity was 50.5%,88.2%,80.6%,84.9%,81.7%,respectively.In the verification set,the AUC was 0.731,0.724,0.713,0.780,0.799,the accuracy was 73.2%,70.7%,68.3%,73.1%,78.0%,and the sensitivity was 76.2%,61.9%,88.1%,76.2%,78.6%,respectively.The specificity was 70.0%,80.0%,47.5%,70.0%and 77.5%,respectively.By DeLong's test,in the training set there were statistically significant differences between combined model and the clinical model,the intratumoral model and the peritumoral model(Z=3.660,2.791,2.201,P=0.0003,0.005,0.028).There was no significant difference between the combined model and the intratumoral+peritumoral model(Z=1.583,P=0.114).The results showed that the combined model in the training set and validation set was better than the clinical model,intratumoral model,peritumoral model and intratumoral+peritumoral model in predicting the status of HER-2.DCA showed that the combined model had higher clinical utility than the clinical model,intratumoral model,peritumoral model and intratumoral+peritumoral model at risk thresholds of 13%-60%in the training set.Conclusions:The combined model based on DCE-MRI intratumoral and peritumoral radiomics combined with clinical and imaging features can better predict the expression status of HER-2 in breast cancer patients.